Can't make the plugin to full edit multiple files with same prompt

[stoperro]

The plugin works great overall for me (for completion and editing), but I'm struggling with one scenario.

I have a prompt I use to edit whole files (by selecting whole file content and Ctrl+Shift+E) and it works suprisingly well. Now, the project has ~200 files and I wouldn't want to do it for each file one by one, each time waiting for AI to finish. I would like it to process all files while I do something else, then just fix up a few mistakes it makes in all files.

There doesn't seem to be an option to apply edit to multiple files at once, which would solve my problem (I would be fine with applying them on disk, I can correct mistakes refering to diff of course).

Thus, I've tried editing with Agent with no success. It does long analysis of what it needs to do, which usually results in him telling me there's too much files, or giving me useless script to do the edits without AI (the changes I apply are not super complex, but simple regex is not enough...).

I've tried reproducing it on smaller project with simpler prompt, still no success. Only once it actually wrote something to disk (created useless script in .go to write hardcoded comment).

To reproduce, download https://github.com/go-chi/chi, and apply agent edit:

For all .go files in the project, simply edit them with below prompt:

For functions that don't have documentation comment, add one line summary comment just before the function.

I might be really bad at using this tool/agent edit, though in reality I really don't want to even use agent, just mass apply of full file edit for each file (e.g. for files selected in file view in vs code, or better. the search where I could search for files I would like to edit with regex).

[igardev]

@stoperro Interesting use case! I didn't predict it. For now it is not possible to edit all files with one command, but I will think about it. Will write here if I implement it.

As for the agent - a good tools model could do the changes in several files. You could try attaching (with @) a file (or 2 files) and ask something like "For all attached files, for functions that don't have documentation comment, add one line summary comment just before the function.". However, it will not work well with many attached files (context limits).

[stoperro]

Thank you. I needed the feature for what I'm doing right now and decided to give it a try to modify llama.vscode to do that, and it seems to work. I put PR with my changes, maybe it would be useful at least as a reference.

I thought it's possible to get file list from search result, but couldn't find the way, but I have no idea what I was doing honestly (luckily AI helped). So as a workaround, I added a query for glob pattern for project files, which was good enough for my case, but it's far from perfect.

The AI edited 176 files over 4 hours (used 14B model on my 2x3090), made some smaller or bigger mistakes and now I have a working copy changes that I need to go through. This will still save me a few days of very tedious manual edits :)

[igardev]

@stoperro Good go hear that llama.vscode + your improvements saves you days of tedious work. Thanks for the pull request, I will review it.

[igardev]

@stoperro I've merged your pull request. Will add some small things (confirmation before starting editing, new item in the llama-vscode menu, etc.) and will include it in the next version. Thank you for the contribution!

[stoperro]

I did some more research on the topic. I noticed that some large files weren't edited properly likely due to context size. Thus I tried figure out a way for llama.vscode to not only search for files, but inside the files to search for exact functions to be edited, so that I can limit what's edited, and not process whole file.

But before that I've noticed llama.vscode actually don't provide context to edit text selection (like lines before and lines after), and tried to fix that, but failed miserably while trying to create the right prompt that will provide that additional information to chat model. Likely the model I used (Qwen2.5 Instruct Coder) wasn't trained on the data similar to my prompt. I realized that even if I've succeeded, likely it would have worked for that particular model only and if user has changed the model, AI would again start misbehaving. So even if I've made some additional regex input, likely results would be bad anyway.

So I've forfeit that, and decided stick to full file editing and instead to switch to model with larger context, seems Qwen3 has 2x the context of 2.5 so I've used that. Unfortunately they provide only 30B and 480B, but I've found some 4bit quantization which barely fit on my 2*24GB VRAM. I've also increased Chats_max_tokens option to 256000 just in case. I've used "asdgad/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M-GGUF" as didn't find smaller Qwen3 Coder Instruct on ggml org. With that I was able to process larger files, which was taking long for them and maybe wasting some CPU cycles on unrelated code, but that was fine for me.

Bottom line: so far, for mass file editing context size is king.